{
  "paper_id": "12f10cc21b869a508d42bc231aafdfc2",
  "paper_title": "Enhancing Indoor Robot Pedestrian Detection Using Improved PIXOR Backbone and Gaussian Heatmap Regression in 3D LiDAR Point Clouds",
  "paper_authors": [
    "Duy Anh Nguyen",
    "Khang Nguyen Hoang",
    "Nguyen Trung Nguyen",
    "Duy Anh Nguyen",
    "Hoang Ngoc Tran"
  ],
  "publication_date": "2024-01-09",
  "publication_venue": "IEEE Access",
  "abstract": "Accurate and robust pedestrian detection is fundamental for indoor robotic systems to navigate safely and seamlessly alongside humans in spatially constrained, unpredictable indoor environments. This paper presents a novel method, IRBGHR-PIXOR, a detection framework specifically engineered for pedestrian perception in indoor mobile robots. This novel approach employs an enhanced adaptation of the cutting-edge PIXOR model, integrating two pivotal augmentations: a remodeled convolutional backbone leveraging Inverted Residual Blocks (IRB) in unison with Gaussian Heatmap Regression (GHR), as well as a Modified Focal Loss (MFL) function to tackle data imbalance issues. The IRB component notably bolsters the network's aptitude for processing intricate spatial representations generated from sparse 3D LiDAR scans. Meanwhile, integrating GHR further elevates accuracy by enabling precise localization of pedestrian subjects. This is achieved by modeling the probability distribution and predicting the central location of individuals in the point cloud data. Extensively evaluated on the large-scale JRDB dataset comprising intense scans from 16-beam Velodyne LiDAR sensors, IRBGHR-PIXOR accomplishes exceptional results, attaining 97.17% Average Precision (AP) at the 0.5 IOU threshold. Notably, this level of accuracy is achieved without significantly increasing model complexity. By enhancing algorithms to overcome challenges in confined indoor environments, this research paves the way for safe and effective deployment of autonomous technologies once encumbered by perceptual limitations in human-centered spaces. Nonetheless, evaluating performance in diverse edge cases and integration with complementary sensory cues promise continued progress. The developments contribute towards the vital capacity of reliable dynamic perception for next-generation robotic systems coexisting in human-centric environments.\n",
  "keywords": [
    "Robot navigation",
    "pedestrian detection",
    "pedestrian tracking",
    "PIXOR",
    "Gaussian heatmap",
    "point cloud",
    "deep learning"
  ],
  "paper_institutions": [
    "Ho Chi Minh City University of Technology (HCMUT), Vietnam National University Ho Chi Minh City",
    "FPT University, Can Tho"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "deepseek-v4-flash",
  "prompt_template": "simple",
  "extracted_at": "2026-08-18T10:30:31.168430Z",
  "template_tag": "simple",
  "summary": "### 摘要\n\n本文针对室内移动机器人在空间受限、光照变化、传感器噪声和密集人群等复杂条件下行人检测精度不足的问题，提出了一种基于改进 PIXOR 的 3D LiDAR 行人检测框架 IRBGHR-PIXOR。相关工作表明，PIXOR 等单阶段检测器在 KITTI 等室外场景表现优异，但直接用于 JRDB 这类室内数据集时面临小目标、类别不平衡和复杂背景等挑战。该方案在 PIXOR 骨干中用倒残差块（IRB）替代瓶颈块，结合跳跃连接与转置卷积实现多尺度特征融合，并在检测头引入高斯热图回归（GHR），将行人中心建模为关键点，核心机制为 K(x,y)=exp(-((x-px)^2+(y-py)^2)/(2σp^2))，同时提出修正焦点损失（MFL）以缓解数据不平衡并强调精确定位。实验采用 JRDB 数据集的 16 线 Velodyne LiDAR 点云，共 15,000 个样本按 80/10/10 划分训练/验证/测试，使用旋转、缩放和平移增强使样本翻倍至 30,000；输入为 0.05m 分辨率的 BEV 表示，尺寸 240×240×30，检测范围 x/y 为 [-6,6]m、高度 [-1.5,1.5]m，模型从零训练，采用 AdamW 优化器、one-cycle 策略、学习率 3e-4、batch size 10 和 20 epochs。在 JRDB 测试集 0-6m 范围内，IRBGHR-PIXOR 在 IoU=0.5 时达到 97.17% AP，优于 PIXOR 和 AFDet；消融实验显示骨干改进带来约 6.5% AP 提升，高斯热图带来约 3-4% AP 提升，修正焦点损失与 Smooth L1 组合在验证集达到 97.21% AP@IoU=0.5 和 88.30% AP@IoU=0.5:0.9。该方法适用于咖啡馆、商场、大厅等室内真实场景，但在 6-24m 远距离和高 IoU 阈值下与 AFDet 差距小于 1%。未来可探索 GAN 数据增强、自适应损失调优、多数据集评估、LiDAR 与 RGB/深度传感器融合以及在线域适应，以进一步提升鲁棒性和泛化能力。\n\n### 关键要点\n\n- 提出 IRBGHR-PIXOR 框架，在 PIXOR 骨干中用倒残差块替换瓶颈块，使计算量降低约一个数量级，并增强多尺度特征提取能力。\n- 在检测头引入高斯热图回归，将行人中心作为关键点，用二维高斯分布生成软标签，以提升小尺寸行人的定位精度。\n- 提出修正焦点损失（MFL），对远离中心的地面真值像素施加更大惩罚，从而缓解类别不平衡并强调精确行人定位。\n- 采用鸟瞰图（BEV）表示 LiDAR 点云，以 0.05m 分辨率生成 240×240×30 的输入，检测范围为 x/y 轴 [-6,6]m、高度 [-1.5,1.5]m。\n- 在 JRDB 数据集上评估，该数据集包含 15,000 个样本，按 80/10/10 划分训练、验证和测试集，数据来自咖啡馆、商场、礼堂等多种室内外地点。\n- 数据增强采用旋转（±20 度）、缩放（0.95-1.05 倍）和平移（0.4 倍），使训练样本翻倍至 30,000，并将验证集 AP@IoU=0.5 从 91.06% 提升到 97.21%。\n- 模型使用 PyTorch 从零训练，采用 AdamW 优化器、one-cycle 策略、学习率 3e-4、batch size 10、20 epochs，训练环境为 Ubuntu 20.04、Intel i7、GTX 3060 和 32GB RAM。\n- 在 JRDB 测试集 0-6m 范围，IRBGHR-PIXOR 在 IoU=0.5 时达到 97.17% AP，在 IoU=0.7 和 0.9 时分别为 95.27% 和 77.48%，优于 PIXOR 和 AFDet。\n- 与原始 PIXOR 相比，骨干网络改进带来约 6.5% AP 提升；加入高斯热图回归后 AP 提升约 3-4%，证明其对小目标和复杂背景的有效性。\n- 损失函数消融显示，Modified-Focal 与 Smooth L1 组合最佳，在验证集 0-6m 达到 97.21% AP@IoU=0.5 和 88.30% AP@IoU=0.5:0.9。\n- 参数调优表明，修正焦点损失中 α=2、β=4 时性能最高，且训练收敛更快，最小损失约 0.5，而 α=1、β=4 时约为 1.2。\n- 局限与未来方向：在 6-24m 远距离且 IoU=0.9 时比 AFDet 低不到 1%；未来可探索 GAN 增强、元学习自适应损失、多数据集验证、多模态融合和在线域适应。",
  "available_templates": [
    "deep_read",
    "simple"
  ]
}